Professional Documents
Culture Documents
Recombinant Uncertainty in Technological Search
Recombinant Uncertainty in Technological Search
Technological Search
Lee Fleming
Harvard University, Graduate School of Business, Morgan Hall T97,
Boston, Massachusetts 02163
lfleming@hbs.edu
mercial application or adoption of an invention. from Kroeber’s (1948) image of a tree of cultural
He argued that, “the making of the invention and artifacts. Unlike genetic trees, however, the branches
the carrying out of the corresponding innovation are, of the tree of technology can fuse together. “Separate
economically and sociologically, two entirely differ- types or branches fuse together to produce new types,
ent things” (p. 85). While these sources remain inter- which merge once again with still other branches”
dependent, their conflation confuses our explanations (Kroeber 1948, p. 138). Similar to the theory of natural
of their respective processes and contributions to evolution before the discovery of DNA, the analogy
uncertainty. To avoid such confusion, this paper will breaks down at the genetic level. “We who postulate
focus on invention and purely technological sources theories of techological evolution likewise have our
of uncertainty. Darwins but not our Mendels” (Kroeber 1948, p. 210).
Attempts to understand the source of technological Although the idea of recombination as the source of
uncertainty have also been frustrated by a lack of novelty has been widely discussed, the implications of
understanding of the process by which inventors the idea remain undeveloped. Previous work suggests,
create new technologies. A theory of the process however (Schumpeter 1939, Henderson and Clark
of invention would facilitate our understanding of 1990), that an invention can be defined as either a
the sources of technological uncertainty. To address new combination of components or a new relationship
this issue, this paper synthesizes two classic perspec- between previously combined components. While
tives on the sources of technological novelty: first, knowledge, science, algorithms, culture, applications,
invention is a process of recombination, and, second, and manufacturing processes also influence invention,
invention is indeed an inherently uncertain and, there- they are not part of an actual artifact and are not
fore, typically local search process. actually instantiated in an invention. They strongly
influence the process of inventive search, however,
by inspiring, aiding, explaining, or constraining the
Invention As a Process of use of particular components or combinations. Collo-
Recombinant Search quial usage of the word “components” implies known,
Many scholars have proposed that recombination pro- available, and commercially available hardware. In
vides the ultimate source of novelty (Gilfillan 1935, this paper, however, “components” will denote the
Usher 1954). Schumpeter (1939, p. 88) observed that constituents of invention, along the lines of what
“innovation combines components in a new way, Schumpeter calls “factors” (1939, p. 88).
or that it consists in carrying out New Combina-
tions.” Nelson and Winter (1982, p. 130) state that The Scope of Potential Recombination
“the creation of any sort of novelty in art, science, Even though the number of potential components
or practical life—consists to a substantial extent of overwhelms the imagination, there are no restrictions
a recombination of conceptual and physical materi- on the scope of their recombination. Components are
als that were previously in existence.” Henderson and not like genes and “similar” technologies are not
Clark (1990) argue that the mere rearrangement of like species (Basalla 1988). In contrast to variation
previously used components can itself cause destabi- processes within genetically isolated populations,
lizing industrial change. Hargadon and Sutton (1997)
inventors can recombine any components within their
describe how a design consulting firm creates nov-
purview. Perceptions that certain technologies or com-
elty by brokering ideas and technologies between their
ponents “belong together” develop through social
clients’ industries.1 Basalla (1988) develops an analogy
construction and previous association. For example,
if an electrical engineer of the 1940s had been asked
1
While the concept of recombinant search as the source of novelty
about his profession’s use of sand and aluminum,
may well apply outside the strictly focused realm of invention—
indeed, to cuisine, innovation (Abernathy and Utterback 1978, he probably would have replied with a blank stare.
Kogut and Zander 1992, Levinthal 1998), and process improvement Today, he or she probably would reply that they are
(Romer 1993)—I can only comment empirically on invention. the most common basic materials of semiconductors
and the focus of much research investment. Clearly, no by cognitive and social phenomena. The most
technology evolves independently of the entire world fundamental influence is a limitation on the num-
of made things. At any point in technological evolu- ber of potential components and combinations that an
tion, any component is at risk of being recombined inventor can simultaneously consider. Because every
with any other component. The made world evolves invention can be incorporated in further recombina-
as a holistic, continuous, and interdependent web, and tions, inventors’ combinatoric potential has grown
not as a disjoint assortment of separate trajectories or explosively (Weitzman 1996). The set of potential com-
product life cycles. binations and, a fortiori, the possible ways that each
Investors constantly import previously untried set of potential combinations can be combined has
components from outside the extant made world, become essentially infinite. It has become impossible
for example, the use of medicinal substances from for individual inventors, groups, even entire commu-
tropical jungles or the exploitation of petroleum and nities of inventors to have more than an infinitesimal
natural gas, substances that had little application to understanding of all these potential combinations
technology prior to the 19th century. Investors also and relationships. As a result of this combinatoric
create new components through encapsulation and explosion, inventors and their organizations and
hierarchical modularization of component sets (Simon communities must focus and recombine locally from
1996, Baldwin and Clark 1999). Such “black box” engi- a limited set of components and combinations.
neering efforts make the underlying components less These arguments follow the assumption of bounded
salient but do not destroy the recombinant potential rationality and local search (March and Simon 1958,
of the components or the encapsulating module. As Nelson and Winter 1982, Cohen and Levinthal 1990,
these processes make additional components available March 1991, Kauffman 1993, Stuart and Podolny
for further recombination, the diversity of the “made” 1996). Localness corresponds to inventors’ familiar-
world increases. Because all previously used compo- ity with their recombinant search space. Local search
nents and inventions provide potential constituents or exploitation (March 1991) occurs when an inven-
for future inventions, the potential risk set for recom- tor recombines from a familiar set of technology
bination is the entire extant made world. Because components or refines a previously used combina-
inventors also scan outside the made world, anything tion. A pastry chef searches extremely locally when
not yet derived from the natural world can also be he mixes previously used dyes in a new propor-
considered as a potential constituent of invention, as tion to create a novel frosting color. Distant search
part of the theoretical risk set. Recombination usually or exploration (March 1991) occurs in the opposite sit-
occurs, however, between components that are salient, uation, when inventors try completely new compo-
proximal, and available for the inventor. nents or combinations. The early auto industry pro-
These ideas prompt us to look backward in time vides many examples of successful and unsuccessful
to consider the components and untried combinations distant search including various power sources, pneu-
that were available at the time of invention; they matic tires, brakes for each passenger, four-wheeled
prompt us to look forward to predict which inven- diamond configurations, and, in the 1930s, a combined
tions are more likely to motivate further recombi- car and plane configuration (Basalla 1988).
nation. Prediction of further recombination depends One would expect that local recombination is more
on the fundamental tension between exploration of certain and, on average, more successful. To the
untried possibilities and exploitation of previous suc- extent that inventors draw from familiar compo-
cesses (March 1991). nent sets and refine previous combinations, they are
less likely to develop a completely useless invention.
Cognitive, Social, and Technological They also decrease their upside potential, however,
Influences on Recombination of developing a radically different invention that is
Because the agents of recombination are people, of much greater impact. Local recombination is more
the process of invention remains strongly influenced certain because inventors learn from past failures.
Inventors learn which components failed in previ- Parallel arguments hold for the refinement and
ous inventions and stop using them. They learn to improvement of combinations. Engineers gain expe-
avoid the combinations and architectures that failed rience with particular combinations by using them.
in the past. They winnow and bound the less suc- As inventors learn which combination relationships
cessful regions of recombinant space (Vincenti 1990). or architectures are less useful, they avoid those
This winnowing and bounding improves the average approaches. Such learning and knowledge helps
usefulness of inventive efforts but also decreases the them improve their inventive efforts on average and
possibility of wildly successful inventions. decrease inventive uncertainty. As with knowledge
As inventors reuse components they begin to about individual components, knowledge about com-
understand and characterize those components. They bination relationships increases with frequency of use
begin to understand which components are more and decreases since last use. As with individual com-
or less useful in different contexts. Such knowledge ponents, knowledge of combinations and their opti-
enables selection and exploitation of more appropriate mal relationships will diffuse throughout the made
components in future inventions. Understanding and world. These arguments imply combination hypothe-
knowledge increase with use; the greater the use of a ses that parallel the individual component hypotheses.
component, the greater the knowledge of and famil- Hypothesis 3. Refinement of familiar combinations
iarity with it. Knowledge can also be forgotten, how- will increase an invention’s usefulness.
ever, and is therefore more potent the more recently it
has been gained (Argote et al. 1990). More recent and Hypothesis 4. Refinement of familiar combinations
frequent usage therefore implies greater knowledge will decrease inventive uncertainty.
and familiarity. Inventions that incorporate familiar Practitioners have long recognized the value of
components should be more useful because inventors reuse and refinement (Mead and Carver 1980) and the
can select more appropriate components. Because they difficulty of exploring new regions of the essentially
can better predict the performance of the included infinite design space. For example, Altschuler (1998)
components, incorporation of familiar components explicitly recommends searching previous inventions
should also decrease inventive uncertainty and, hence, for universal analogies and possible applications to
the variability of outcomes. new contexts, and Goldenberg et al. (1999) propose
Inventors can draw on others’ knowledge and and empirically validate a method to identify and
experience in addition to their own. While social prox- link profitable dependencies between previously used
imity certainly increases the ease and likelihood of components. Unfortunately for inventors, however,
sharing, knowledge about component use will still dif- these benefits of familiarity do not last forever. This
fuse between organizations (Allen 1977) and techno- results from the technological and social-psychological
logical communities (Bijker 1987). Many mechanisms exhaustion of potential refinements, given a particular
facilitate knowledge diffusion, including person- combination. Both influences run counter to the posi-
nel movement, personal friendships, organizational tive effects of local search.
merger, education, reverse engineering, technical liter- Technological exhaustion occurs because most of
ature, and strategic alliances (Ahuja 2000). Although the possible relationships between a set of compo-
much knowledge will be lost or changed in its dif- nents have already been tried. As Kim and Kogut
fusion, sharing will take place at all social levels, (1996) argue, “The repeated application of a particular
between individuals, organizations, and communities. set of technologies or organizing principles eventually
Taken together, these arguments imply the following. exhausts the set of potential combinations.” The argu-
ment generalizes Sahal’s (1985) demonstration of the
Hypothesis 1. Recombination of familiar components
decreasing returns to physical scaling. For example, if
will increase an invention’s usefulness.
semiconductor inventors restricted their usage to their
Hypothesis 2. Recombination of familiar components original materials of aluminum and bipolar transis-
will decrease inventive uncertainty. tors, progress in the field would have halted long ago.
Citations dependent invention’s usefulness or importance prior art citations by future patents to focal patent
Mean technology control expected citations to technically similar patents weighted fixed effects by focal patent’s sub-class
membership in class
Variance technology control expected variance in citations to technically similar weighted fixed effects by focal patent’s sub-class
patents membership in class
Number of prior art citations control patents that cite more heavily should be more number of prior art citations by focal patent
heavily cited
Single-class dummy control process of recombination is finer grained than can equals 1 if focal patent assigned to only one subclass
measure
Number of subclasses control number of invention’s components number of focal patent’s subclasses
Newest subclass control artifact of patent classification system equals the minimum number of previous uses
amongst the focal patent’s sub classes
Number of classes control breadth of patent classifications number of focal patent’s classes
Component familiarity independent inventor’s familiarity with components of the recent and frequent usage of focal patent’s subclasses
invention across all U.S. patents
Combination familiarity independent inventor’s familiarity with particular combination recent and frequent usage of particular subclass com-
of components bination across all U.S. patents
Cumulative combination usage independent cumulative number of inventive trials with exactly cumulative number of U.S. patents with exactly same
same combination subclass combination
Bibliometric studies have repeatedly demonstrated historical consistency in the measurement of compo-
that future prior citations to a patent correlate with nents across time.
its technological importance and value (Albert et al. I do not propose that inventors recombine patent
1991, Hall et al. 2000). To make maximal use of the subclasses directly, only that subclasses can be used
data, I measure prior art citations to a focal patent for to observe indirectly the process of recombinant
6 years and 5 months after its granting. This period search and learning. I will illustrate the correspon-
should capture the bulk of citations to a patent as these dence between components and patent subclasses
citations typically reach a plateau after about three with patent 5,136,185, coauthored with John Walther.
years from the grant date (Jaffe et al. 1993). While there It was (as of Dec. 1996) classified in four subclasses:
appears to be strong correlation between the rates of 326/16 (with test facilitating feature), 326/31 (signal
early and later citations to a patent, researchers are sensitivity or transmission integrity), 326/56 (tristate
actively pursuing the topic (Hall et al. 2000). (i.e., high impedance as third state)), and 326/82
To measure the independent variables, I proxy com- (current driving (e.g., fan in/out, off chip driving,
ponents with patent subclasses. The patent office cat- etc.)). Each of these subclasses corresponded to a
egorizes all patentable technologies into some 400 well-understood digital hardware component at that
“class references”. Each class is also subdivided into time. Each of them can be found in (and, for this
very fine divisions of technology or approximately patent, were drawn specifically from) contemporary
100,000 “subclasses” (Trajtenberg et al. 1997). The textbooks of digital design such as McCluskey (1986).
patent office typically assigns each patent into mul- A test facilitate feature (p. 426) has become a neces-
tiple subclasses within and across major classes. The sity with large computer chips due to their millions
patent office also establishes and updates new classes of gates. To test the innards of a chip, the chip must
and subclasses each year, as technology changes (Carr be operated in a “test mode.” A transmission gate
1995, p. 128). This retrospective updating enables (p. 118) simply passes on a signal when enabled.
A tristate driver (back cover diagram of digital com- Average component familiarity of patent
ponents and pp. 105, 119, 142) is a gate that can drive a
all subclasses Iij
j of patent i
bus or be turned off to present high impedance to the i s subclasses ≡ Fi = (2)
all subclasses 1
bus. Finally, “fan in” and “fan out” (p. 104) refers to j of patent i
the number of components that can drive or be driven
I set the time constant of knowledge loss at five
by a particular gate (high fan out basically means a
years. This constant in the denominator of the expo-
big gate). While we were not aware of the subclass
nential implies that approximately one-third of the
definitions at the time of the invention, we were very
knowledge remains after five years, or a yearly loss
familiar with these basic building blocks of our tech-
rate of 18%. Argote, Epple, and Darr (Argote et al.
nological community, namely digital hardware engi- 1990, Epple et al. 1991, Darr et al. 1995) have estimated
neers. I propose that subclasses serve as proxies for a much higher geometric loss parameter for manu-
these building blocks. facturing and service organizations, between 40% to
With the subclassifications and date of application 97% per year. It is unlikely that design technology
of each patent (interpolated prior to 1975), I develop would experience such a high rate of loss, however.
the three independent variables for each focal patent: Design knowledge is far less contextual and more eas-
component familiarity, combination familiarity, and ily articulated than manufacturing or service experi-
cumulative combination usage. Component familiarity ence. Design knowledge is more likely to have been
proxies inventors’ familiarity with components, based recorded in trade journals, firm documentation and,
on the average degree to which the components have of course, patents. It may have been actually real-
been recently and frequently used. The assumption ized in prototypes and products and is far more
is that inventors will be more familiar with compo- likely to have required substantial personal effort and
nents that have been recently and frequently used. investment on the part of its designers. All of these
I first calculate an individual measure for each sepa- influences would argue for a slower loss rate than
rate subclass of the focal patent. To the extent that a manufacturing and service organizations, but the pre-
particular subclass has been recently and frequently cise estimation certainly constitutes a valid research
used, its individual measure will be higher. For each question.
focal patent, I look backward to 1790 and consider I calculate combination familiarity and cumulative
each of its individual subclasses in turn. Whenever combination usage similarly. The cumulative usage
measures how many times since 1790 a particular
a particular subclass has been used in any previous
combination of subclasses has been used. Combina-
invention, I multiply the indicator of occurrence by
tion familiarity proxies inventors’ familiarity with the
an exponentially decaying component (1). This expo-
combination, based on the degree to which the com-
nential component represents the loss and forgetting
bination has been recently and frequently used. For
of knowledge. For example, it is more likely that an
each focal patent, I consider its particular combina-
inventor will have learned from previous use of a
tion of subclasses. I then look back in time at all
subclass, if that sub-class was used three years prior, other patents that used an identical combination of
instead of thirty. These occurrences are then summed subclasses. For each previous patent that used an
and averaged (2). identical combination of subclasses, I increment the
Individual component familiarity of patent i’s cumulative combination usage (3). For the combi-
subclass j ≡ nation familiarity variable, I multiply each indicator
count by the exponential component, to reflect the
Iij = 1 patent k uses subclass j (1) loss and forgetting of experiential knowledge regard-
all patents k granted
before patent i ing the combination (4). All the measures remain indi-
application date of patent i− application date of patent k rect because they assume that learning occurs with use
−
×e time constant of knowledge loss and that knowledge diffuses throughout technological
communities and the made world. Given that all the extent that industries have similar technologies and
independent variables are highly skewed, I took the citation patterns.
square root of their original value. This minimized Average citations to a patent classified within class
the effects of outliers and enabled more parsimonious i ≡ i =
modeling.
1patent j cites a patent within class i
all patents i granted
from 1985−19905
∗proportion of cited patent within class i
Cumulative combination usage of patent i ≡ Ci =
1patent k classified within class i
all patents k granted
∗proportion of patent k within class i
1patent k uses identical (3) from 1985−1990
all patents k granted
before patent i (5)
combination of subclasses
as patent i Variance in citations to patents classified within class
i ≡ i2 =
Combination familiarity for patent i ≡ Ri = 1patent k classified within class i ∗ proportion
all patents k granted
from 1985−19905
of patent k within class i ∗ citesk − i 2
1patent k uses identical combination
all patents k granted
1patent k classified within class i
before patent i all patents k granted
of subclasses as patent i from 1985−19905 ∗proportion of patent k within class i
application date of patent i − application date of patent k
(6)
×e time constant of knowledge loss
Table 2 Descriptive Statistics for All U.S. Patents, May/June 1990 ties of maximum likelihood estimators (Greene 1993,
n = 17264
p. 133).
Variable mean stnd.dev. minimum maximum
Table 3 Correlation Matrix for All U.S. Patents, May/June 1990 n = 17 264
cites mean variance prior single numsub newest class comp combin
variable decreases inventive uncertainty, its effect 1/i . Integrating (10) with the density function of (11)
on the variance of this error distribution should be gives the probability of the negative binomial of (12).
negative.
Most derivations of the negative binomial start from i i i−1 −
i i
g
i =
e for i > 0
a basic Poisson model (7). The basic Poisson model
i i
estimates the probability of an observed count, condi- and
= t −1 e−t dt (11)
0
tional on an expected mean i . To avoid negative (i.e., i yi
undefined) expected values for the mean i , Poisson
yi + i i i
Pryi xi = (12)
models typically parameterize explanatory variables yi !
i i + i i + i
as an exponential function (8). The method of maxi-
While the first term of (9) fully specifies the mean
mum likelihood is then applied to the joint frequency
of the negative binomial, various parameterizations
formed from the product of the marginal frequencies
of i remain possible. Cameron and Trivedi (1986)
of (7), to determine the coefficient values that are most
propose the negative binomial II parameterization (or
likely to result in the observed counts.
Negbin II model) when the variance/mean ratio of
e−i i i
y the observed data is linear in the mean. By contrast,
Pryi xi = (7) the Negbin I holds the variance/mean ratio constant.
yi !
The Negbin II specification is also much more robust
Eyi xi = i = exi (8) to distributional misspecification than other param-
eterizations (Cameron and Trivedi 1986). I verified
The negative binomial model replaces the Poisson
the applicability of the Negbin II by regressing pre-
mean i with the random variable ˜ i (9). This replace-
dicted counts on the quantity (residualsˆ2/predicted).
ment enables the inclusion of an error term
i = ei —
The coefficient and intercept were positive thus sup-
and allows the predicted mean to vary according to
porting the Negbin II parameterization (Cameron and
the distribution of the error term. Substitution of ˜ i
Trivedi 1986). Also consistent with Cameron and
for i in (7) results in (10).
Trivedi’s argument (1986), Negbin II models demon-
˜ i = exi +i = i
i (9) strated much more significant log likelihoods than
Negbin I models.
e−i
i i
i yi
Pryi xi
i = (10) Equation (13) specifies the Negbin II parameteriza-
yi ! tion of the conditional variance. i is the inverse of
When ˜ i replaces i in (10), the probability of the i and is parameterized as an exponential function,
observed count becomes conditional on the error dis- similar to the mean specification but with potentially
tribution. This conditioning can be removed, however, different variables. Since var
i = 1/i = i , greater
by specifying the error distribution and integrating variance of the error term for a given estimated mean
with its probability density function to obtain the will result in an increase of i . I operationalize uncer-
marginal density. Most formulations specify a gamma tainty by estimating the effects of the causal vari-
distribution for
i with parameter i and probabil- ables on the dispersion parameter . Variables that
ity density function g
i as in (11) (Hausman et al. decrease will decrease the variability and hence the
1984, Cameron and Trivedi 1986, King 1989, Long uncertainty of inventive outcomes.
1997). Although the error term can take other distribu-
i
tions, this parameterization is flexible, computation- Varyi x = 1 + = i + 2i (13)
i
ally tractable, and can be derived from a variety of
assumptions. While other versions of the gamma dis- STATA estimates (12) by the method of maximum
tribution take two parameters, this derivation (from likelihood. This technique estimates both the mean i
Long 1997, p. 232) sets both to i , which forces the and variance 1/i from (12), and hence produces only
mean of
i equal to one and the variance of
i equal to a single log likelihood.
Table 4 Negative Binomial Models of Citation Counts (All U.S. Patents, May/June of 1990)
results on the dispersion. The negative effect supports it would be difficult to compile similarly vast observa-
Hypothesis 4, that inventive uncertainty decreases tions across so many communities. The dataset there-
with the refinement of previously used combinations. fore trades detail and depth in exchange for breadth
At the mean citation count of 3.80 and the mean value and possibility of observation.
of combination familiarity at 0.28, the variance is 2.2% In addition to differences in the dependent variable
less than it would have been without the effect of com- of citations across communities, the accuracy of the
bination familiarity on the dispersion. At the maxi- independent variables also varies across communities.
mum value of combination familiarity, the difference Subclasses and combinations of subclasses represent
grows to 64%. only a proxy for inventors’ components and architec-
Finally, Hypothesis 5 predicts that cumulative tures. While the subclasses of digital hardware patents
use of a particular combination will eventually correspond very closely to my engineering experi-
exhaust recombinant potential. The results support ence, other subclasses may not. For example, financial
this hypothesis, as indicated by the negative and sig- patents tend to be classified in fewer subclasses and,
nificant estimate on cumulative combination usage. hence, may not reflect a process of recombinant search.
When cumulative combination count reaches its max- They may also be better characterized as knowledge or
imum value of 26.76, a patent receives only 37.9% of algorithms and not technology. Such patents, however,
the citations it would have received at the variable’s became popular only after the observation period.
lowest value of 0. A one standard deviation increase of In addition to concerns about data and variables,
cumulative combination use results in a 5.6% decrease these results also remain open to alternate interpreta-
in expected citations. tions. Most importantly, these models cannot defini-
tively separate learning and familiarity, technological
exhaustion, and life-cycle effects. For example: an
Discussion exogenous variable (such as the inventive myopia
These results should be viewed cautiously for a described by Henderson (1995)) could drive the use
variety of reasons. The typical reservations regard- of particular components and combinations, such that
ing the use of patent data certainly apply, most increased citations might simply reflect the popular
notably that patenting practices and effectiveness vary usage of those components and combinations. As a
across industries (Levin et al. 1987). Furthermore, even result, the positive signs in the mean for compo-
though these data cover all patented technologies nent and combination familiarity and negative sign
across a two-month time period, much inventive activ- for cumulative use might reflect only the normal pro-
ity remains unpatented and, hence, unobserved. These gression of technological life cycles, instead of learn-
issues cause at least three problems. First, the dataset ing and exhaustion. Even with the multiple controls
misses unimportant inventions that failed to merit a for differences in citation patterns across technologies,
patent. Use of citation data mitigates this problem, types of combinations, and retrospective identification
however, because it includes the bulk of relatively use- of seminal subclasses, the models cannot convincingly
less inventions that receive no or few citations. Sec- reject the alternative argument that the results merely
ond, the dataset may miss breakthroughs that firms reflect the normal life-cycle progression.
chose not to patent, presumably for strategic rea- These reservations do not apply, however, to the
sons. However, unless there is systematic bias in those uncertainty hypotheses. Indeed, all of the explana-
firms’ use of particular components or combinations, tory variables in the mean could simply be interpreted
these results should remain valid. Third, technolog- as control variables for the dispersion estimates. The
ical communities vary in their propensity to patent. combination result supports the intuitive argument
Again, while these models controlled for much vari- that uncertainty decreases with refinement and high-
ance across technologies, they did not introduce lights the importance of early architectural refine-
explicit industrial controls. But given the focus on ments as a source of destabilizing technologies for
recombinant search across technological communities, organizations and industries (Henderson and Clark
1990). In contrast, the nonmonotonic results for com- also provide a strong basis for further investigations
ponents does not support the simple linear relation- into the sources of invention and technological uncer-
ship predicted in Hypothesis 2. It remains consistent, tainty. Most importantly, negative binomial count and
however, with the argument that technological break- dispersion models enable researchers to analyze the
throughs derive from new combinations of well-used first and second moments of patent citation data. Such
components (Usher 1954, Nelson and Winter 1982, models will enable us to move beyond basic counts
Sahal 1985). For example, Utterback (1996, p. xxvii) in analysis of patent data and quantitatively ana-
argues that “radical innovations often are seen to lyze the outliers of the highly skewed distributions
be based on the synthesis of well-known technical of inventive trials. Such tools can enable more formal
information or components.” These results also indi- analysis of breakthrough inventions, heretofore, “the
cate that inventors take more inventive risk with domain of economic historians” (Scherer and Harhoff
extremely familiar components. These results support 2000). For example, the classic controversy about the
Utterback’s argument and motivate further research. sources of technological breakthroughs, whether they
For example, are breakthroughs most likely when emerge from smaller, entrepreneurial, and “outside”
inventors combine very familiar components in new firms (Schumpeter 1939, Marquis 1969, Klein 1977), or
combinations? If this were the case, then break- large, industrial incumbents (Schumpeter 1942), can
throughs would be most likely to emerge from social be reconsidered with these data and methods. The
contexts that brought together inventors with deep approach presented here also has application beyond
experience in previously disparate fields. Such con- the study of technology and patents, mainly that we
texts would also be more likely to be the technologi- should think of varying variance as an opportunity
cal source of potential future product life cycles and instead of a nuisance.
trajectories. Future work should also investigate the relation-
ships between invention as a recombinant search pro-
Conclusion cess and other literatures such as integrality, coupling,
This paper developed and tested an explanation for and modularity (Ulrich 1995), complex systems
the sources of purely technological uncertainty. It (Kauffman 1993), modular operators (Baldwin and
argued that the source of technological novelty and Clark 1999, Goldenberg et al. 1999), the evolution of
uncertainty lies within the combination of new com- modular design choices (Simon 1996), and the market
ponents and new configurations of previously com- implications of such issues (Christensen and Verlinden
bined components. Inventors’ experimentation with 2000). For example, interdependence between compo-
new components and combinations leads to less suc- nents should increase inventive uncertainty and mod-
cess on average, but it also increases the variability ularity should decrease it. Kauffman’s (1993) models
that can lead to breakthroughs. Empirical results sup- of search over interdependent landscapes also imply
ported the arguments with the exception that the use a positive but nonmonotonic effect on the mean. Con-
of more familiar components has a nonmonotonic and ceptualizing technological invention as search over
eventually positive effect on the uncertainty of inven- interdependent landscapes implies a complexity catas-
tion. In contrast to the nonmonotonic effect of com- trophe in technological evolution, as inventors face
ponent familiarity, the refinement of previously used greater numbers of components and greater interde-
combinations has a negative and monotonic effect on pendence between them (Fleming and Sorenson 2001).
uncertainty. Inventors are not blind search agents, however, and
Coupled with recent and complementary research their search strategies will differ greatly and pre-
regarding market influences (Adner and Levinthal sumably be more effective than genetic recombination
2000) and formal economic models (Klepper 1996), (Gavetti and Levinthal 2000, Rivkin 2000).
this work helps us understand the causal forces that The work has strategic implications as well. Organi-
underlie the widely observed regularities of the prod- zations that seek technological breakthroughs should
uct life cycle. The data and methods presented here experiment with new combinations, possibly with
old components. They do so, however, at the risk Allen, T. 1977. Managing the Flow of Technology. MIT Press,
of an increased number of failures. If invention is Cambridge, MA.
indeed a partially blind search process, such fail- Altschuler, G. 1998. The Innovation Algorithm. Technical Innovation
Center, Inc., Worcester, MA.
ures are to some extent unavoidable. Simple port-
Amato, I. 1999. Industrializing the search for new drugs. Fortune
folio strategies are unfortunately not the complete
(May 10)
answer, because of the extreme skew of inventive Anderson, P., M. Tushman. 1990. Technological discontinuities and
distributions. Although the overall variance of such dominant designs: A cyclical model of technological change.
distributions decreases with an increasing number Admin. Sci. Quart. 35 604–633.
of trials, firms—and indeed, as Scherer and Harhoff Argote, L., S. Beckman, D. Epple. 1990. The persistence and trans-
(2000) demonstrate—entire economies cannot antici- fer of learning in industrial settings. Management Sci. 36(2)
pate completely stable returns. There might exist orga- 140–154.
nizational mechanisms that encourage recombinant Ayres, R. 1988. Barriers and breakthroughs: An “expanding fron-
tiers” model of the technology-industry life-cycle. Technovation.
exploration, while limiting the downside of increased
7 87–115.
failures. For example, firms that can screen or test
Baldwin, C., K. Clark. 1999. Design Rules: The Power of Modularity.
their nascent inventions more effectively will bene- MIT Press, Cambridge, MA.
fit more and should increase the variability of their Basalla, G. 1988. The Evolution of Technology. Cambridge University
inventive trials. The recent progress in drug design as Press, Cambridge, MA.
a result of automated screening processes provides a Bijker, W. 1987. The social construction of bakelite: Toward a theory
salient example (Amato 1999). Finally, science should of invention. Bijker, Hughes, Pinch, eds. The Social Construction
enlighten and shorten technological search over diffi- of Technological Systems. MIT Press, Cambridge, MA.
cult landscapes. Science can either motivate recombi- Cameron, A., P. Trivedi. 1986. Econometric models based on count
data: Comparisons and applications of some estimators and
nant search across particular technological landscapes,
tests. J. Appl. Econometrics. 1 29–53.
or aid in search across landscapes that inventors have
, F. Windmeijer. 1996. R-squared measures for count data
discovered empirically. regression models with applications to health-care utilization.
J. Bus. Econom. Statist. 14(2) 209–220.
Acknowledgments Carr, F. 1995. Patents Handbook. McFarland, Jefferson, NC.
The author would like to acknowledge the helpful comments of Christensen, C., M. Verlinden. 2000. Disruption, disintegration,
Gautam Ahuja, Steve Barley, Bill Barnett, Phil Bromiley, Laurie and the dissipation of differentiability. Working paper 00-074,
Calhoun, Mike Hannan and his seminar students, James March, Harvard Business School, Boston, MA.
Candace Fleming, Olav Sorenson, Toby Stuart, and three very help- Clark, K. 1985. The interaction of design hierarchies and market
ful reviewers. He would also like to acknowledge the support of
concepts in technological evolution. Res. Policy 14 235–251.
William Simpson and the Faculty Research Computing Center at
Cohen, W., D. Levinthal. 1990. Absorptive capacity: A new per-
the Harvard Business School, the Department of Industrial Engi-
spective on learning and innovation. Admin. Sci. Quart. 35
neering at Stanford University, and Corey Billington, Phil Johnson,
128–152.
and Ellen King and the library staff of Hewlett Packard Company.
Errors and omissions remain my responsibility. Darr, E., L. Argote, D. Epple. 1995. The acquisition, transfer, and
depreciation of knowledge in service organizations: Produc-
tivity in franchises. Management Sci. 41(11) 1750–1762.
Epple, D., L. Argote, R. Devadas. 1991. Organizational learn-
References
ing curves: A method for investigating intra-plant transfer of
Abernathy, W., J. Utterback 1978. Patterns of industrial innovation.
knowledge acquired through learning by doing. Organ. Sci.
Tech. Rev. (June) 40–47.
2(1) 58–70.
Adner, R., D. Levinthal. 2000. Technology evolution and
market heterogeneity: Implications for product and process Fleming, L., O. Sorenson. 2001. Technology as a complex adaptive
innovation. Working paper, INSEAD, France. system: Evidence from patent data. Res. Policy Forthcoming.
Ahuja, G. 2000. Collaboration networks, structural holes, and inno- Gavetti, G., D. Levinthal. 2000. Looking forward and backward:
vation: A longitudinal study. Admin. Sci. Quart. 45 425–455. Cognitive and experiential search. Admin. Sci. Quart. 45
Albert, M. B., F. Narin, D. Avery, P. McAllister. 1991. Direct valida- 113–137.
tion of citation counts as indicators of industrially important Gilfillan, S. 1935. Inventing the Ship. Follett Publishing Co.,
patents. Res. Policy 20 251–259. Chicago, IL.
Greene, W. 1993. Econometric Analysis. Prentice Hall Inc., Upper March, J. 1991. Exploration and exploitation in organizational
Saddle River, NJ. learning. Organ. Sci., 2 71–87.
Goldenberg, J., D. Mazursky, S. Solomon. 1999. Toward identifying , H. Simon. 1958. Organizations. Blackwell Publishers,
the inventive templates of new products: A channeled ideation Cambridge, MA.
approach. J. Marketing Res., 36 200–210. Marquis, D. 1969. The anatomy of successful innovations. Innova-
Hall, B., A. Jaffe, M. Trajtenberg. 2000. Market value and patent tion 1 35–48.
citations: A first look. NBER Working paper 7741, McCluskey, E. 1986. Logic Design Principles with Emphasis on Testable
Hargadon, A., R. Sutton. 1997. Technology brokering and inno- Semi-Custom Circuits. Prentice Hall, Englewood Cliffs, NJ.
vation in a product design firm. Admin. Sci. Quart. 42 Mead, L. Carver. 1980. Introduction to VLSI Systems. Addison-
716–749. Wesley, Reading, MA.
Hausman, J., B. Hall, Z. Griliches. 1984. Econometric models for MicroPatent. 1996. Product of micropatent USA, New Haven, CT.
count data with an application to the patents-R&D relation- Mokyr, J. 1990. The Lever of Riches: Technological Creativity and Eco-
ship. Econometrica 52 909–938. nomic Progress. Oxford University Press, New York.
Henderson, R. 1995. Of life cycles real and imaginary: The unex- Nelson, R., S. Winter. 1982. An Evolutionary Theory of Economic
pectedly long old age of optical lithography. Res. Policy 24 Change. Belknap Press, Cambridge, MA.
631–643. Rivkin, J. 2000. Imitation of complex strategies. Management Sci. 46
, K. Clark. 1990. Architectural innovation: The reconfigura- 824–844.
tion of existing product technologies and failure of established Romer, P. 1993. Ideas and things: The concept of production is
firms. Admin. Sci. Quart. 35 9–30. being re-tooled. The Economist (Sept 11) 70–72.
Jaffe, A., M. Trajtenberg, R. Henderson. 1993. Geographic localiza- Rosenberg, N. 1996. Uncertainty and technological change.
tion of knowledge spillovers as evidenced by patent citations. R. Landau, R. Taylor, G. Wright, eds. The Mosaic of Economic
Quart. J. Econom. 108 577–598. Growth. Stanford University Press, Stanford, CA.
Jorgensen, B. 1997. The Theory of Dispersion Models. Chapman and Ruttan, V. 1959. Usher and Schumpeter on invention, innovation,
Hall, London, U.K. and technological change. Quart. J. Econom. (November)
Kauffman, S. 1993. The Origins of Order. Oxford University Press, Sahal, D. 1985. Technological guideposts and innovation avenues.
New York. Res. Policy 14 61–82.
Kim, D., B. Kogut. 1996. Technological platforms and diversifica- Scherer, F., D. Harhoff. 2000. Technology policy for a world of
tion. Organ. Sci., 7 283–301. skew-distributed outcomes. Res. Policy 29 559–566.
King, G. 1989. Event count models for international relations: Schumpeter, J. 1939. Business Cycles. McGraw-Hill Book Company,
Generalizations and applications. Internat. Stud. Quart. 33 Inc., New York.
123–147. , 1942. Capitalism, Socialism, and Democracy. Harper and Row
Klein, B. 1977. Dynamic Economics. Harvard University Press, Publishers, New York.
Cambridge, MA. Simon, H. 1996. The Sciences of the Artificial. MIT Press, Cambridge,
Klepper, S. 1996. Entry, exit, and innovation over the product life- MA.
cycle. Amer. Econom. Rev. 86(3) 562–583. Stuart, T., J. Podolny. 1996. Local search and the evolution of tech-
, 1997. Industry life cycles. Indust. Corporate Change 6(1) nological capabilities. Strategic Management J. (Summer)
145–181. Tratjenberg, M., R. Henderson, A. Jaffe. 1997. University versus
Kogut, B., U. Zander. 1992. Knowledge of the firm, combinative corporate patents: A window on the basicness of invention.
capabilities, and the replication of technology. Organ. Sci. 3 Econom. Innovation New Tech. 5 19–50.
383–397. Tushman M., P. Anderson. 1986. Technological discontinuities and
Kroeber, A. 1948. Anthropology. Harcourt Brace Jovanovich, Inc., organizational environments; Admin. Sci. Quart. 31 439–465.
New York. Ulrich, K. 1995. The role of product architecture in the manufac-
Levin, R., A. Klevorick, R. Nelson, S. Winter. 1987. Appropriating turing firm. Res. Policy 24 419–440.
the returns from industrial research and development: Com- Usher, A.P. 1954. A History of Mechanical Invention. Cambridge, MA.
ments and discussion. Brookings Paper Econom. Activity (3) 783– Utterback, J. 1996. Mastering the Dynamics of Innovation. Harvard
831. Business School Press, Boston, MA.
Levinthal, D. 1998. The slow pace of rapid technoligical change: Vincenti, W. 1990. What Engineers Know, and How They Know It.
Gradualism and punctuation in technological change. Indust. Johns Hopkins University Press, Baltimore, MD.
Corporate Change 7(2) 217–247. Weitzman, M. 1996. Hybridizing growth theory. Proc. Amer. Econom.
Long, J. 1997. Regression Models for Categorical and Limited Dependent Assoc. (May) 207–212.
Variables. Sage Publications, London, U.K.
Accepted by Karl Ulrich; received October 27, 1998. This paper was with the author 8 months for 3 revisions.